And Another Thing...
An excellent accidental discovery today which has solved an issue I was scratching my head over in terms of odds fetching from api-sports (I thought it maybe a them problem, but it was me!). My fetch from api-sports was often taking the first acceptable bookmaker over/under market from API-Sports, which meant later books with richer O3.5/O4.5 lines were being ignored. I fixed it by ranking valid full-time totals markets by completeness instead of first match, and by accepting alternate market names like Total Goals. Result: better band odds coverage feeding into my ecosystem. Very important as I'm researching goal band cascade forecasting right now :)
6 Jun 2026, 10:33 (UTC)
Perhaps...just perhaps it might be good if the developers of coding agents did something for me. Here is a quote from this morning, as I debug Bet In Play Timer (a research app I'm running to surface signals for betting during play on over goal markets): "I’m wiring the backfill as a proper archive sync, not a one-off hack: the app will reconstruct missing research records from committed picks, then immediately settle any of those whose forecasts are already reconciled. That should make the research archive catch up to the real selected history." - GPT 5.4 "I’m wiring the backfill as a proper archive sync, not a one-off hack" - like what the fcuk. A one off hack was an option? 0_0 Seriously. You see a lot of phrases like that...."I'll inspect the file fully rather than just hand-waving" is another great one. Please. Charge me for those tokens and then go and smoke a bowl GPT. Nice. lol.
5 Jun 2026, 09:32 (UTC)
I've been discussing StatStrike and GoalLab with GPT this morning and received a compliment from the chatbot (not news in itself...validation machines...) that I appreciated: "You seem to care a lot about honest forecasting. That's relatively rare in this space and probably worth leaning into." So let's lean into that as we try to solve the riddle of the churn...25% retention isn't bad...but surely there's room for improvement...
3 Jun 2026, 09:51 (UTC)
Day by day...carefully to ensure I do not carry technical debt forward...Master Feeder is being developed. It will act as the central hub for API calling in my ever expanding forecasting/research ecosystem. It's fun to build :)
21 May 2026, 12:27 (UTC)
Okay. Today I will mostly be working on "Master Feeder" - an API fetching application for macOS that will centralise the API fetch for 8 projects that call the service every day. I have a large daily quota of API calls (150,000 daily atm - highest use was 405,000 when I was fetching three years of global football to train a ML forecaster), but even with that I'm often pushing it in a 24 hour cycle. So I am building the central hub that will fetch for all applications, depending on their needs, and distribute - there's a lot of overlap (daily fixture lists, score polling etc). But ProphIt is a beast. It feeds and feeds. Let's ensure it doesn't go hungry with the Master Feeder.
2 May 2026, 11:37 (UTC)
Bet in Play Timer is almost at 10,000 settled forward tested fixtures. The idea is to identify fixtures where the teams historically have slow starts and then explosive scoring patterns. To see if the patterns are predictive...
30 Apr 2026, 08:55 (UTC)
The first week of V2 of an experiment to determine if sacrificing coverage in large fixture days = positive lift in overall win rate? will complete today. Encouraging early signals. Net lift is 1.3pp. The trim in upload is made by requiring a higher min win rate than usual for the league a fixture is in. V1 was entirely shadow. I was so encouraged that for V2 I allowed the forecaster to turn the rule on for production, and I wasn't wrong for the weekend just passed. What can we squeeze out of the model that uploads for StatStrike iOS app? I'm always looking. Other shadow experiments are running to do with aggregate odds, criteria mixes and so on. My suspicion is I will find, over time, that it is the right move to improve model performance on large fixture days. I suspect users of StatStrike will be happy to have fewer fixtures to scroll on a Saturday if they know what they are seeing has a stronger probability of winning? Clearly having a list predominately hard favs is a ROI risk, but given the hot streak potential of the model, accumulators are a go-go for the canny user. I suspect V2 will remain in production on large volume fixture days. Time will tell. Cheers!
28 Apr 2026, 08:56 (UTC)
I have several research applications running largely automated, gathering archives for analysis. I just looked at one of them and found something very interesting (more than one example in archive of this type now). It's a screenshot of the criteria mix (app displays criteria) and forecaster confidence (100%). it seems very successful with a tally of FT matches that is becoming reliable. More consideration needed. The app is maturing...the 80% is the win rate for this mix at that confidence and the mix seems to highlight high scoring games - 71 is the count of games that went over 4 goals...let's dig deeper.
27 Apr 2026, 13:38 (UTC)
This morning I will mostly be thinking about how best to advise a user of StatStrike how to understand the success of the app. It has very good hot streaks. So I must determine the best advice...ie, if the app has been winning consistently in a long run of fixtures, is it best not to take the next forecasts in the list? Until the app has recorded losses that balance out to the expected daily W/L? ie there has been some regression? Is the pattern consistent enough to be useful?
27 Apr 2026, 11:12 (UTC)
This a test post of the new micro-blog function for all that those random thoughts, and a short form diary. So let's include an attractive looking Spanish Segunda fixture coming up tonight to really test it. See how it copes :)
26 Apr 2026, 17:56 (UTC)